Learning a Terrain- and Robot-Aware Dynamics Model for Autonomous Mobile Robot Navigation

TL;DR

The paper addresses autonomous navigation under variations in terrain friction and robot properties. It introduces TRADYN, a probabilistic forward dynamics model conditioned on a latent robot context $\boldsymbol{\beta}$ and terrain features $\boldsymbol{\tau}$, trained with an ELBO objective and calibrated online to a target environment. Key contributions include explicit joint modeling of terrain- and robot-specific properties, integration of terrain lookups during planning, and demonstration of improved long-horizon prediction and planning efficiency in a 2D unicycle simulation, including robustness analyses under action and observation noise. The work offers a practical, data-driven path to adaptive, energy-efficient navigation in heterogeneous terrains, with potential extensions to real-world scenarios involving partial observability and map learning.

Abstract

Mobile robots should be capable of planning cost-efficient paths for autonomous navigation. Typically, the terrain and robot properties are subject to variations. For instance, properties of the terrain such as friction may vary across different locations. Also, properties of the robot may change such as payloads or wear and tear, e.g., causing changing actuator gains or joint friction. Autonomous navigation approaches should thus be able to adapt to such variations. In this article, we propose a novel approach for learning a probabilistic, terrain- and robot-aware forward dynamics model (TRADYN) which can adapt to such variations and demonstrate its use for navigation. Our learning approach extends recent advances in meta-learning forward dynamics models based on Neural Processes for mobile robot navigation. We evaluate our method in simulation for 2D navigation of a robot with uni-cycle dynamics with varying properties on terrain with spatially varying friction coefficients. In our experiments, we demonstrate that TRADYN has lower prediction error over long time horizons than model ablations which do not adapt to robot or terrain variations. We also evaluate our model for navigation planning in a model-predictive control framework and under various sources of noise. We demonstrate that our approach yields improved performance in planning control-efficient paths by taking robot and terrain properties into account.

Figures (13)

• Figure 1: Terrain- and robot-aware robot dynamics models for control-efficient navigation. We propose a novel approach for learning dynamics models for control-cost optimal navigation. Our learned dynamics model can adapt to unobserved robot properties, such as the mass, and properties of the terrain, such as the friction coefficient which can vary spatially. The above example shows planned paths to two different goals (white/black crosses: start/goal). Our method avoids areas of high friction coefficient and favors areas of low friction coefficient. As the impact of the friction coefficient depends on the mass of the robot, a heavy robot ($m= 4kg$, blue, orange) takes longer detours to the goal than a light robot ($m= 1kg$, green, red).
• Figure 2: Terrain- and robot-aware forward dynamics model (TRADYN). We embed the initial robot state $\bm{x}_0$ as hidden state of a gated recurrent unit (GRU). The GRU predicts the next state in the latent space for which it receives context encoding $\bm{\beta}$ and embeddings of action $\bm{u}$ and terrain observation $\bm{\tau}$ as additional inputs. Latent states are decoded into Gaussian distributions on the robot's next state. While during training the actual terrain observation ${\tau}(\bm{x}_n)$ is used, during prediction, the map ${\tau}$ is queried at predicted robot locations ${\tau}({\color{red}\bm{\hat{x}}_n})$. See \ref{['sec:tradyn:method']} for details.
• Figure 3: Exemplary rollouts (length 50) on two different terrain layouts (rows) and for two exemplary robot configurations (low-inertia, high-inertia) (columns). Rollouts start from the center; actions are sampled time-correlated. The low-inertia robot has minimal mass $m=1$ and maximal control gains $k_\mathrm{throttle}=1000$, $k_\mathrm{steer}=\pi/4$. The high-inertia robot has maximal mass $m=4$ and minimal control gains $k_\mathrm{throttle}=500$, $k_\mathrm{steer}=\pi/8$. Equally colored trajectories (, , ) correspond to identical sequences of applied actions. See \ref{['sec:tradyn:simenv']} for details.
• Figure 4: Relationship of RGB terrain features $\bm{\tau}$ (left column) to friction coefficient $\mu$ (right column). See \ref{['sec:tradyn:terrainlayouts']} for details.
• Figure 5: Prediction error evaluation for the proposed model and its ablations (no terrain lookup / no calibration), plotted over the prediction horizon (number of prediction steps). From left to right: Positional error (euclidean distance), velocity error (absolute difference), angular error (absolute difference). Depicted are the mean and 20%, 80% percentiles over 150 evaluation rollouts for 5 independently trained models per model variant. Our approach with terrain lookup and calibration clearly outperforms the other variants in position and velocity prediction (left and center panel). For predicting the angle (right panel), terrain friction is not relevant, which is why the terrain lookup brings close to no advantage. However, calibration is important for accurate angle prediction. See \ref{['sec:tradyn:eval_prediction']} for details.